Short answer
Regulation (EU) 2024/1689, the EU Artificial Intelligence Act (AI Act), does not put every product into one of four mutually exclusive boxes. It applies several layers. Article 5 prohibits specified practices, so no operator may lawfully place, put into service or use an AI system for those purposes except within an express exception. Articles 6 and Annexes I and III identify high-risk AI systems; their providers carry the main design, documentation, conformity-assessment and post-market duties, while importers, distributors and deployers have distinct checks and use obligations. Article 50 imposes targeted transparency duties on providers or deployers, depending on who can technically mark the output or communicate with the affected person.
Chapter V separately regulates providers of general-purpose AI (GPAI) models, with extra duties for models presenting systemic risk. A downstream product built on a GPAI model can also be high-risk, prohibited, or subject to Article 50. The model provider and the downstream system provider therefore may each have duties at different points in the value chain.
Timing matters. The original prohibitions have applied since 2 February 2025, and the GPAI rules since 2 August 2025. Article 50 generally applies from 2 August 2026. Regulation (EU) 2026/1744 (the Digital Omnibus on AI), effective on 27 July 2026, moved the relevant Annex III high-risk rules to 2 December 2027 and the Annex I product high-risk rules to 2 August 2028. Its two new intimate-content prohibitions apply from 2 December 2026.
Start with the actor, intended purpose and conduct
The provider develops, or has developed, an AI system or GPAI model and markets it or puts it into service under its own name. A deployer uses an AI system under its authority, except personally and non-professionally. Importers and distributors place or make systems available through the EU supply chain.
Classification turns on intended purpose and conduct, not a marketing name. The same model might support a low-risk writing tool, a high-risk recruitment system and a prohibited manipulative practice. Article 2 can also reach non-EU actors that market in the EU or whose system output is used there.
Article 4, as replaced by the Digital Omnibus, requires providers and deployers to take measures supporting the development of AI literacy among staff and others operating systems on their behalf, in light of their knowledge, training, context and affected people. It expressly does not require a guaranteed level for any individual. Other applicable EU law continues to operate alongside the AI Act.
Prohibited practices: the activity itself is the problem
Article 5 is conduct-based. Its already-applicable prohibitions include materially harmful manipulation or exploitation of vulnerability; specified social scoring; individual criminal-risk assessment based solely on profiling or personality traits; untargeted scraping to build facial-recognition databases; emotion inference in workplaces or schools, subject to medical and safety exceptions; specified biometric categorisation; and most real-time remote biometric identification in public spaces for law enforcement.
“The following AI practices shall be prohibited:”
—
Each prohibition has precise elements. For example, Article 5(1)(a) requires a specified technique, material distortion of informed decision-making and significant harm or its reasonable likelihood. The elements and exceptions must be tested, not inferred from a label.
From 2 December 2026, new points (ba) and (bb) also prohibit specified AI generation or manipulation of non-consensual intimate material and child-sexual-abuse material. Provider-side liability is limited to systems intended for that result, or systems where it is a reasonably foreseeable and reproducible outcome without reasonable and adequate safeguards. Deployer-side liability requires purposeful use to generate or manipulate the material. Article 5(1a) and (1b) contain further conditions and limits.
Responsibility follows the prohibited act. A provider or importer may breach Article 5 by placing a system for a prohibited purpose on the market or putting it into service; a deployer may breach it through prohibited use. The narrow exceptions for real-time remote biometric identification in law enforcement impose safeguards on the deploying authority, including necessity and proportionality, impact assessment, registration, ordinarily prior authorisation, notification and reporting. They do not excuse other prohibited practices.
High-risk systems: classification triggers a shared compliance chain
Article 6 supplies two routes. First, an AI system is high-risk if it is a safety component of, or itself is, a product covered by Annex I harmonisation law and that product requires third-party conformity assessment. The amended definition confines a safety function to an intended purpose of preventing or mitigating health-and-safety risks; mere integration in a regulated product is insufficient. Article 6(1a)–(1c) further exclude specified non-safety functions and third-party assessments triggered only by non-health-and-safety risks. Second, Annex III lists use cases such as certain biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration and administration of justice.
“In addition to the high-risk AI systems referred to in paragraph 1, AI systems referred to in Annex III shall be considered to be high-risk.”
—
An Annex III system is exceptionally not high-risk where it does not pose a significant risk to health, safety or fundamental rights, including because it does not materially influence decision-making, and it satisfies an Article 6(3) condition such as performing a narrow procedural or preparatory task. Profiling natural persons always defeats that exception. A provider relying on the exception must document its assessment and register the system.
When the relevant rules begin applying—2 December 2027 for the Annex III route and 2 August 2028 for the Annex I route—the duties divide by role:
- Provider: meet the lifecycle requirements, complete conformity assessment, declaration, CE marking and registration, monitor the system, correct non-compliance and report serious incidents.
- Authorised representative: perform mandated tasks for a non-EU provider and hold specified documentation for authorities.
- Importer and distributor: perform their respective pre-market checks, preserve compliance and stop or recall non-conforming systems.
- Deployer: follow instructions, provide competent human oversight, monitor use, retain controlled logs and act on risks. Specified deployers must also complete a fundamental-rights impact assessment.
Contracts cannot defeat the statutory allocation. Under Article 25, rebranding, substantial modification or repurposing can make an importer, distributor, deployer or third party the provider. A product manufacturer can similarly become provider when the high-risk system is placed on the market or put into service with its product under its name.
For some Article 6(1) systems, delegated acts may eventually limit specified Articles 9–15 and 17–25 duties where Annex I sector law gives equivalent or higher protection and the AI Act's overall protection is not reduced. That mechanism does not itself erase duties before the Commission specifies its scope.
Article 50 transparency duties: responsibility follows who can disclose
Article 50 is narrower than the high-risk regime and can apply even where a system is not high-risk. It allocates duties according to practical control:
- The provider of a system interacting directly with people must inform them that it is AI, unless that is obvious in context.
- The provider of a system generating synthetic content must make outputs machine-readable and detectable as artificial, subject to stated exceptions.
- The deployer must inform people exposed to emotion recognition or biometric categorisation.
- The deployer must disclose deepfakes and specified AI-generated public-interest text, subject to the Act's qualifications.
“The information referred to in paragraphs 1 to 4 shall be provided to the natural persons concerned in a clear and distinguishable manner at the latest at the time of the first interaction or exposure.”
—
Provider marking does not remove deployer disclosure or high-risk duties.
Article 50 generally applies from 2 August 2026. A targeted transition gives providers of synthetic-content systems placed on the market before that date until 2 December 2026 to comply with Article 50(2)'s machine-readable marking duty. The Omnibus changed the code-of-practice mechanism in Article 50(7), not the basic allocation in paragraphs 1–5.
General-purpose AI: regulate the model and its downstream integration separately
A GPAI model displays significant generality, performs a wide range of tasks and can be integrated into varied downstream systems. Its provider must maintain technical documentation, give downstream providers compliance information, adopt an EU-copyright policy and publish a sufficiently detailed training-content summary. A limited open-source exception removes the first two duties for qualifying models but does not cover systemic-risk models.
A third-country GPAI provider generally appoints an EU representative. High-impact capability or Commission designation can make a model systemic-risk; training compute above 10²⁵ floating-point operations creates a presumption. The provider then has evaluation, systemic-risk mitigation, incident-reporting and cybersecurity duties.
The key distinction is between model and system. The GPAI-model provider bears Chapter V model duties. A downstream business integrating that model into an AI system is the provider of that system and must classify its intended use. If the resulting system is high-risk or falls under Article 50, downstream obligations arise as well; if it is used for an Article 5 practice, the prohibition controls.
A practical classification sequence
- Confirm the covered system, actor and territory.
- Identify each operator and any rebranding, modification or repurposing.
- Test Article 5, both Article 6 routes and the Article 6(3) exception.
- Test provider and deployer duties under Article 50 independently.
- Allocate Chapter V model duties separately from downstream-system duties.
- Apply the correct date to each resulting obligation.
Conclusion
The AI Act uses cumulative layers. Prohibited-practice duties attach to the operator that markets, commissions or carries out the forbidden conduct. High-risk compliance is provider-led but shared across the supply chain. Article 50 splits transparency between the provider that can design or mark the system and the deployer controlling the encounter or publication. GPAI obligations govern the model provider, while downstream providers retain their classification and compliance duties. A reliable assessment maps actor, purpose, value-chain changes and application date before assigning a label.
Sources
- — checked for interpreting judgments; none was identified for this question as of 29 July 2026.
This article provides general legal information, not advice for a specific matter. Laws, facts and procedural deadlines may change the result. Follow the linked primary sources and obtain local advice before acting.